نوع مقاله : مقاله پژوهشی
نویسندگان
دانشکده مهندسی نقشه برداری و اطلاعات مکانی، دانشکدگان فنی، دانشگاه تهران
چکیده
آتشسوزیهای گسترده جنگلی با تخریب پوشش گیاهی، تشدید ناپایداری خاک و تغییر کارکرد زیستبومها، به یکی از جدیترین چالشهای محیطی عصر حاضر تبدیل شدهاند و شناسایی دقیق نواحی سوخته پس از حریق، پیشنیاز ارزیابی خسارت، برنامهریزی احیا و مدیریت خطر است. دادههای اپتیکی ماهوارهای، بهویژه تصاویر سنتینل-۲، همراه با شاخصهای طیفی پرکاربرد، ابزار قدرتمندی برای نقشهبرداری مناطق سوخته فراهم میکنند، اما کارایی آنها به انتخاب مناسب شاخصها و مدلهای طبقهبندی وابسته است. هدف این پژوهش، ارزیابی و مقایسه کارایى یک روش آماری کلاسیک، سه الگوریتم یادگیری ماشین و دو معماری یادگیری عمیق در تشخیص نواحی سوخته، بر پایه ترکیب باندهای سنتینل-۲ و شاخصهای طیفی، در آتشسوزی جنگلی ست. نتایج این پژوهش نشان داد تمامی مدلها قادر به تفکیک الگوی کلی سوختگی از پسزمینه نسوخته بودند، اما در دقت عددی و نویز مکانی اختلاف قابل توجهی داشتند. روش آماری مبتنی بر MLE، اگرچه دقتی نزدیک به ۹۸ درصد را بهدست آورد، بهدلیل مقدار بالای طبقهبندی نادرست پیکسلهای نسوخته بهعنوان سوخته و تولید لکههای پراکنده در حاشیه ناحیه سوخته، کمترین قابلیت اتکا را ارائه نمود. در میان الگوریتم RF بهترین عملکرد با دقت ۹۹٫۶۷% را داشت. SVM نیز با F1-score بیش از ۹۶ درصد عملکردی پایدار و رقابتی نشان داد. الگوریتم Adaboost ، با وجود بهبود محسوس نسبت به روش آماری، بهسبب حساسیت به نمونههای دشوار، میزان بالاتری از عدم شناسایی پیکسلهای واقعاً سوخته را ایجاد کرد. دو مدل یادگیری عمیق، یعنی CNN و MLP، نقشههایی پیوسته و کمنویز تولید نمودند و از نظر دقت نتایجی بسیار نزدیک به RF بهدست آوردند.
موضوعات
عنوان مقاله [English]
Comparison of Different Methods for Burned Area Detection Using Spectral Indices and Sentinel-2 satellite imagery: Statistical, Machine Learning, and Deep Learning (A Case Study of the Kenneth Wildfire in Los Angeles)
نویسندگان [English]
- Mahdi Hasanlou
- Zohreh Roodsarabi
- Parvin Hasan Teymori
School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran
چکیده [English]
Background and Objectives: Large-scale wildfires, through the destruction of vegetation, increased soil instability, and disruption of ecosystem functioning, have become one of the most serious environmental challenges of the modern era. Accurate post-fire burn-area delineation is essential for damage assessment, restoration planning, and risk management. Satellite optical data—particularly Sentinel-2 imagery—combined with widely used spectral indices provide a powerful basis for mapping burned areas; however, their performance depends strongly on the choice of indices and classification models. The objective of this study is to evaluate and compare the effectiveness of a classical statistical classifier, three machine learning algorithms, and two deep learning architectures for burned-area detection, using a combination of Sentinel-2 spectral bands and spectral indices in the Kenneth wildfire in Los Angeles.
Methods: Following selection of the post-fire Sentinel-2 imagery and cloud masking, eight core spectral bands (visible, near-infrared, red-edge, and shortwave infrared) along with five commonly used indices related to burn severity, vegetation condition, and moisture content were extracted, forming a 15-variable input image for model development. Binary reference labels (burned/unburned) were derived from the official wildfire incident database, and spatially random sampling was used to create training (70%) and testing (30%) subsets. All features were normalized using min–max scaling. Subsequently, a classical Maximum Likelihood Estimation (MLE) classifier, three machine learning algorithms—Adaptive Boosting (AdaBoost), Random Forest (RF), and Support Vector Machine (SVM)—and two deep learning models—Convolutional Neural Network (CNN) and Multilayer Perceptron (MLP)—were trained. Model evaluation was performed using confusion-matrix metrics including Accuracy, Precision, Recall, F1-score, and Intersection over Union (IoU). Feature importance was also calculated for each algorithm.
Findings: All models successfully distinguished the general burn pattern from the unburned background; however, they differed substantially in numerical accuracy and spatial noise. The MLE classifier, although yielding nearly 98% accuracy, showed the lowest reliability due to a high rate of misclassified unburned pixels (FP) and scattered artifacts around burn perimeters. Among machine learning methods, RF exhibited the best performance, achieving ~99.67% Accuracy, ~97% F1-score, and the highest IoU, with the lowest FP and FN values. SVM also showed stable and competitive performance with an F1-score exceeding 96%, though slightly more boundary-related errors than RF. AdaBoost improved notably over the statistical classifier but, due to sensitivity to difficult samples, produced higher FN values. Both deep learning models (CNN and MLP) generated smooth, low-noise burn maps and achieved Accuracy, F1-score, and IoU values closely matching RF. Feature-importance analysis indicated that shortwave infrared bands (SWIR-1, SWIR-2) and burn/vegetation indices—particularly NBR, NDVI, and SAVI—were the most influential predictors, whereas visible bands contributed less to model decisions.
Conclusion: The results demonstrate that integrating Sentinel-2 infrared bands with vegetation and moisture indices, combined with machine learning and deep learning models, provides an accurate and robust framework for post-fire burn-area mapping in heterogeneous landscapes. RF, followed by CNN and MLP, emerges as the most effective set of models for operational implementation, while MLE and AdaBoost serve better as baseline methods. Key limitations include reliance on a single wildfire event and single-date post-fire data; thus, extending the framework to multiple fire regimes, diverse vegetation types, and more complex topographic conditions, as well as incorporating multitemporal data and radar/altimetry sensors, is recommended for future research. The findings support the development of operational wildfire monitoring systems, prioritization of restoration zones, and sustainable resource-management planning in fire-prone regions.
)